Medical Imaging Data Analysis based on Unsupervised Domain Adaptation for Tubular Structure Segmentation

Abstract

Medical image analysis plays a crucial role in clinical medicine, encompassing preoperative diagnosis of diseases, lesion evaluation, intraoperative clinical treatment, and postoperative recovery assessment. A common challenge in medical imaging datasets is the scarcity of labels and the limited data availability. With the rise of deep learning technologies, methods based on convolutional neural network (CNN) have achieved remarkable success in both segmentation and classification tasks. However, as deep learning technologies evolved, the high demand for large quantities of high-quality labeled data in medical images limited the applicability of fully supervised deep learning techniques. To address this issue and reduce reliance on high-quality labels, unsupervised methods, especially unsupervised domain adaptation (UDA) methods, have been extensively studied. Tubular structures in medical images, such as nerves, blood vessels, and cell membrane boundaries, have been a critical focus of research. Variations in data acquisition methods and instrumentation lead to domain gaps in imaging characteristics even within the same anatomical site, while tubular structures across different anatomical sites exhibit substantial differences in distributions and prior anatomical knowledge. In this thesis, we propose novel UDA methods for medical image segmentation, with a particular focus on the segmentation of tubular structures in medical images. The proposed method addresses key challenges in the analysis of medical image data for tubular structures. Our thesis methods are mainly divided into two parts. The first part of the thesis methodology focuses on the application of UDA in a single anatomical source domain, presenting single-source UDA and multi-source UDA approaches. In the second part of the thesis methodology, we explore the application of UDA in cross-anatomical sites, addressing the unique challenges posed by variations across anatomical sites.